引導式測驗 · 中等 等級
檢查點分片和可恢復訓練測驗
練習檢查點元資料、訓練狀態恢復和重新啟動驗證,包括權重分片和可恢復狀態之間的差異。
問題 1 的 8
為什麼大模型檢查點要分成分片?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 為什麼大模型檢查點要分成分片?
- 讓模型在小輸入上訓練得更快
- 對使用者隱藏模型權重
- 因為分片提高了準確性
- 因為單一檢查點對於一個檔案或機器來說可能太大而無法有效處理
- Which extra state is typically needed for training resumption beyond model weights?
- 僅最終準確率分數
- Optimizer and scheduler state, step progress, RNG state, and relevant data-loader progress
- Only a model-weight shard index
- A screenshot of the training dashboard
- 長期工作可恢復培訓的主要好處是什麼?
- 它消除了對驗證集的需要
- 它保證模型永遠不會過度擬合
- 崩潰或搶佔不會強制從頭開始重新啟動
- 它消除了優化器
- In a sharded distributed checkpoint, how can ranks avoid gathering the entire state on one worker?
- By saving their assigned local tensor shards with checkpoint metadata
- By each emailing a full model copy
- By deleting optimizer state before every save
- By disabling all state serialization
- 索引檔案在分片檢查點中扮演什麼角色?
- 它儲存用戶的密碼
- 它將每個參數名稱映射到包含它的分片文件
- 它壓縮資料集中的影像
- 它安排學習率
- 哪種格式通常用於模型分片的安全性、記憶體映射載入?
- MP3
- CSV
- 安全張量
- 僅 ZIP 存檔
- A package contains safetensors weight shards and an index, but no optimizer state. What can it support by itself?
- Loading model weights, without fully restoring optimizer training history
- Guaranteed exact resumption of every interrupted run
- Recovering the next batch position automatically
- Reconstructing every random-number-generator state
- A sharded checkpoint loader supports a new rank layout. What should the team verify before trusting recovery?
- That fewer ranks guarantee identical floating-point results
- That the weight index alone includes every application state
- That no restart test is necessary
- That the necessary saved state loads correctly and the resumed training step behaves as expected